acceptodds
Under review as a conference paper at ICLR 2027

Time-Adaptive Fractional Spiking Network for Irregular Multivariate Time Series Forecasting

Abstract

Irregular multivariate time series (IMTS) forecasting is challenging due to irregular sampling, variable asynchronicity, and partial observability. Most existing methods for IMTS modeling primarily rely on temporal alignment or structural reformulation to alleviate the computational difficulties caused by irregular sampling. However, these methods often fail to fully exploit the sparse event characteristics arising from the partial observability of IMTS. With event-driven computation and sparse activation, spiking neural networks (SNNs) provide a natural modeling paradigm for preserving the event semantics of IMTS and capturing their temporal dependencies. Thus, we propose the Time-Adaptive Fractional Spiking Network (TAFNet), a multi-granularity framework for IMTS forecasting. We introduce a time-adaptive fractional leaky integrate-and-fire (TAF-LIF) neuron that jointly incorporates actual inter-observation intervals and a fractional memory kernel into the membrane potential evolution, enabling interval-aware temporal modeling and persistent historical memory. We further construct multi-granularity temporal representations to capture scale-dependent dynamics under varying observation densities, and integrate them with global periodic information extracted directly from irregular timestamps. Extensive experiments on real-world datasets demonstrate the state-of-the-art performance of TAFNet.

Primary area:
learning on time series and dynamical systems
Keywords:
Spiking Neural Networks, Irregular Multivariate Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_C5cQOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TAFNet, a spiking framework for irregular multivariate time-series (IMTS) forecasting. Its main component is the Time-Adaptive Fractional LIF (TAF-LIF) neuron, which combines fractional-order membrane dynamics with actual inter-observation intervals. Specifically, TAF-LIF adopts a non-uniform L1 approximation of the Caputo derivative, such that historical membrane increments are weighted according to their actual temporal distances, and further adapts the membrane time factor according to the current inter-observation interval. Based on TAF-LIF, the proposed TAFNet combines an event-aligned spike encoder with multi-granularity temporal patching and frequency-guided fusion using non-uniform Fourier features. Experiments on HumanActivity, USHCN, PhysioNet, and MIMIC show competitive forecasting performance against both ANN and SNN baselines. The paper also provides ablations and controlled comparisons of TAF-LIF with vanilla LIF, f-LIF, and EA-LIF.

Critical strengths

1. The paper addresses a meaningful problem by extending fractional spiking dynamics to irregularly sampled time series.

Instead of treating irregular observations as uniformly spaced discrete steps, TAF-LIF explicitly incorporates actual temporal intervals and uses a non-uniform L1 approximation of the Caputo derivative. Consequently, historical membrane increments are weighted according to their physical temporal distances. I find this formulation well motivated for IMTS and technically reasonable.

2. The experimental evaluation is reasonably comprehensive and includes useful controlled neuron-level comparisons.

The paper evaluates four IMTS benchmarks and compares against both ANN and SNN forecasting methods. Importantly, Table 3 compares TAF-LIF with vanilla LIF, f-LIF, and EA-LIF within the same TAFNet backbone, while keeping the other architectural components and training settings unchanged. This provides a relatively focused evaluation of the proposed neuronal dynamics.

Critical weaknesses

1. The incremental contribution of TAF-LIF over the closely related f-LIF is not sufficiently isolated.

Both approaches employ fractional-order neuronal dynamics, while TAF-LIF appears to introduce two main modifications: (i) non-uniform timestamp-aware fractional weighting and (ii) an interval-adaptive membrane time factor τk\tau_k. However, the current experiments do not disentangle the individual contributions of these two modifications. Therefore, although Table 3 shows that TAF-LIF performs better than f-LIF on average, it remains unclear which component is responsible for this improvement and whether both components are necessary. I consider this an important issue for assessing the methodological contribution beyond existing fractional spiking neurons.

2. The empirical advantage over f-LIF is relatively modest on several datasets, while the evidence supporting persistent fractional historical memory is limited.

For example, on MIMIC, the reported MSE changes only from 0.4246±0.00390.4246 \pm 0.0039 for f-LIF to 0.4237±0.00410.4237 \pm 0.0041 for TAF-LIF. More importantly, the final configuration uses K=2K=2 on all four datasets, meaning that the fractional historical-memory term contains at most one preceding state increment. The paper further reports that increasing the memory length does not monotonically improve performance. These observations make it unclear whether the main empirical benefit comes from persistent fractional memory or primarily from better modeling of irregular physical time intervals.

Questions for authors

1. Can the authors disentangle the two main modifications over f-LIF?

In particular, I would like to understand the performance of: (a) fractional dynamics with the proposed non-uniform timestamp-aware weighting but a fixed membrane time factor, and (b) interval-adaptive τk\tau_k without the proposed non-uniform fractional historical weighting. A controlled comparison among these variants, f-LIF, and the complete TAF-LIF would substantially clarify the source of the reported improvement.

2. Are the improvements of TAF-LIF over f-LIF statistically significant?

Some differences appear small relative to the reported standard deviations. For example, on MIMIC, the MSE is 0.4246±0.00390.4246 \pm 0.0039 for f-LIF and 0.4237±0.00410.4237 \pm 0.0041 for TAF-LIF. Since the experiments are conducted using five random seeds, paired per-seed results or an appropriate statistical test would help establish whether the proposed neuron provides a consistent improvement.

3. Given that K=2K=2 is selected for all four datasets, how should the claimed role of fractional historical memory be interpreted?

With K=2K=2, the historical term contains at most one preceding state increment, and increasing KK does not consistently improve performance. Does this indicate that the primary benefit of TAF-LIF comes from modeling irregular physical time intervals rather than persistent or long-range fractional memory?

The above questions are central to my current assessment rather than suggestions for future work. In particular, convincing evidence that the timestamp-aware fractional formulation and interval adaptation provide complementary and consistent improvements over f-LIF would substantially strengthen the paper, and I would be open to raising my score accordingly.

Reviewer_WrJzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces the Time-Adaptive Fractional Spiking Network for irregular multivariate time series forecasting. The architecture utilizes a time-adaptive fractional leaky integrate-and-fire neuron to process non-uniform sampling intervals and asynchronous time series. The model achieves competitive MAE and MSE on multiple datasets.

Critical strengths

Forecasting Irregular multivariate time series is non-trivial as data points are sampled non-uniformly, and each channel of the time series is async in nature. Most algorithms assume aligned and fixed length input.

This work proposes an interesting idea, by introducing fractional order dynamics, the spiking neuron can utilize both the sampling data points and the interval for encoding. Experiments show this work achieves competitive performance. Spiking neuron with Fractional order dynamics is worthy of more investigation, and may potentially be applied to new applications and improve performance of SNNs.

Critical weaknesses

Though the methodology is interesting, I have following concerns.

  1. Omission of Training Mechanics: Firing and resetting are essential operations in spiking neurons. these operations are non-differentiable. Authors do not mention how they overcome this.

  2. Unclear learnable parameters: Section 4 introduces a "learnable Non-Uniform Discrete Fourier Transform", but does not specify which variables are optimized. Similarly, while \alpha is defined as learnable, its parameter update mechanism is not detailed.

  3. Superficial Visualization: Experiments and results shall provide useful insights or justify claims. But the observation rate versus firing rate visualization in Figure 4 provides little useful information. A positive correlation between input density and spike frequency is an expected behavior, not an insight that justifies the architectural choices.

  4. Unclear computational efficiency profiling: though the computational efficiency analysis in Appendix isolates FLOPs and SOPs, it remains unclear if the NUDFT, the fractional history calculations and other calculations are fully accounted for in the reported FLOPS.

  5. Disjointed origination of section 4: While the authors do include a brief introductory paragraph at the beginning of Section 4 to summarize the pipeline, the actual structural organization of the subsections disrupts the chronological data flow. Subsections are organized topically rather than chronologically (e.g., introducing the global frequency module in 4.1 before defining the primary temporal patching mechanism in 4.2, and main contribution in 4.3). This structural disjointedness makes it hard to understand the full flow, forcing the reader to reconstruct the sequential execution of the pipeline mentally.

Questions for authors

1: how do the parameters such as \alpha and the Non-Uniform Discrete Fourier Transform are trained? 2: how do authors overcome the non-differentiable problem of spiking behavior? 3: what is the exact protocol to calculate flops?

1 comment

Thanks for replying. I don't have additional concerns.

Reviewer_DuR1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TAFNet for irregular multivariate time series forecasting. It introduces time adaptive fractional LIF neuron to capture irregular temporal dependencies and multi granularity modeling and frequency guided fusion.

Critical strengths
  1. The paper addresses a good problem in irregular time series forecasting.

  2. The proposed TAF-LIF neuron provides an interesting approach to modeling temporal dependencies.

Critical weaknesses
  1. The framework uses window-based patching and global frequency modeling, which may require processing the entire observation window rather than incrementally handling asynchronous observations.

  2. The memory length is set to 2. It remains unclear whether such a short memory window is sufficient to capture long-range dependencies and why such small.

Questions for authors
  1. Why is K=2 sufficient for modeling long-range temporal dependencies, and why not select larger?

  2. Can TAFNet process observations incrementally as they arrive at different timestamps, without reprocessing the entire observation window and the result is same?

1 comment

Thank you for the detailed responses. My concerns have been adequately addressed, and I will increase my rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.